feat: add invoke functionality to web-ui

[avi-alpert]

Summary

• Add deployed invocation support to the web UI for all four agent protocols (HTTP, A2A, AGUI, MCP)
• Extract resolveInvokeTarget into a reusable module shared by CLI invoke and web-ui handlers
• Return deploymentTargets from the resources API so agent inspector can build target picker
• Add ?target=deployed path to the MCP proxy (/api/mcp) for deployed MCP tool listing and calling
• Fix mid-stream errors in deployed invocation handlers — emit SSE error frame and always call res.end()

Additions to agent inspector (web-ui) APIs

Deployed invocation (POST /invocations?target=deployed)

• Accepts agentName, targetName, prompt, sessionId, userId in the request body
• Routes to the correct handler based on agent protocol (HTTP/A2A/AGUI)
• Streams responses as SSE events matching the local invocation format

Deployed MCP proxy (POST /api/mcp?target=deployed)

• Supports initialize, tools/list, and tools/call JSON-RPC methods
• Reuses mcpInitSession, mcpListTools, mcpCallTool from the AWS SDK layer
• Returns JSON-RPC envelope matching the local proxy response shape

Resources API (GET /api/resources)

• Returns deploymentTargets: [{ name, region, description? }] from aws-targets.json

Error handling

• All three deployed streaming helpers (HTTP, A2A, AGUI) now wrap for await in try/catch/finally
• On mid-stream failure: writes an SSE error frame, then ends the response (previously left client hanging)

Test plan

• Unit tests for deployed invocation path (validation, protocol routing, mid-stream errors, config failures)
• Existing MCP proxy tests updated and passing
• Manual: agentcore dev → toggle to Deployed mode → invoke HTTP/A2A/AGUI agent
• Manual: agentcore dev → MCP agent → toggle Deployed → verify tools/list and tools/call work
• Manual: verify local invocation still works for all protocols (no regression)

[github-actions]

Package Tarball

aws-agentcore-0.14.1.tgz

How to install

gh release download pr-1326-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.14.1.tgz

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-cli-automation]

Nice refactor extracting resolveInvokeTarget into a reusable module — the unit tests in resolve.test.ts are well-structured and only mock at the I/O boundary (token fetch). The new web-UI deployed-invocation path is a useful addition, but I have a few concerns that should be addressed before merging.

Summary of issues:

1. Mid-stream errors in the new handleDeployed*Invocation helpers leave the client hanging (headers sent, but no res.end() and no error frame).
2. ~250 lines of new handler code in invocations.ts (the entire deployed-invocation path) has zero test coverage.
3. handleDeployedInvocation doesn't accept targetName, so multi-target projects can only invoke against the first deployed target.

Details inline.

[agentcore-cli-automation]

LGTM. Verified the three issues raised in earlier automated review comments have been addressed in the latest commit:

• Mid-stream errors: each handleDeployed*Invocation helper now wraps the for await in try/catch/finally that emits an SSE error frame and ends the response.
• targetName plumbing: the request body now includes targetName and it's forwarded to resolveInvokeTarget.
• Test coverage: deployed-invocations.test.ts covers the validation paths, protocol routing (HTTP/A2A/AGUI), the resolve-error path, the mid-stream failure path, and the config-load failure path.

The resolveInvokeTarget extraction looks clean, the existing handleInvoke is faithfully refactored onto it, and resolve.test.ts gives good coverage of the auth/session/baggage edge cases.

No new blocking concerns from me.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-cli-automation]

Earlier review concerns (mid-stream errors, targetName plumbing, handleDeployedInvocation test coverage, handleDeployedMcpProxy test coverage) have all been addressed in the latest commits — nice work. I do see two new shape/inconsistency issues in the deployed MCP proxy that I think should be sorted out before merging, plus an FYI on telemetry. Details inline.

[agentcore-cli-automation]

FYI / non-blocking: deployed invocation paths have no telemetry.

The equivalent agentcore invoke command emits cli.command_run with protocol/streaming attrs (src/cli/commands/dev/command.tsx:257), and the telemetry README explicitly calls out instrumenting new features. The new deployed-invocation paths in the web UI (and the deployed MCP proxy in mcp-proxy.ts) don't emit anything.

I realize the local web-ui invocation paths above also lack telemetry, so this is partly a pre-existing gap rather than a regression — flagging in case you want to address both in this PR or a follow-up. Up to you whether to block on this; if you're going to defer it, an issue/TODO would be helpful so we don't lose track.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-cli-automation]

All serious issues from the prior review rounds have been addressed:

• Mid-stream errors silently dropped — fixed: each deployed streaming helper now wraps the for await in try/catch/finally, emits an SSE error frame, and always calls res.end() (invocations.ts:515-523, 553-561, 593-601).
• targetName not plumbed through — fixed: parsed from request body and forwarded to resolveInvokeTarget in both invocations.ts:373-379, 422 and mcp-proxy.ts:107-122, 156.
• Test coverage for deployed-invocation path — added: __tests__/deployed-invocations.test.ts covers validation, protocol routing (HTTP/A2A/AGUI), mid-stream failure, resolve errors, and config load failure.
• Test coverage for handleDeployedMcpProxy — added: __tests__/mcp-proxy.test.ts ?target=deployed describe block covers all 8 listed cases.
• mcpListTools leaking mcpSessionId — fixed: const { tools } = await mcpListTools(...) (mcp-proxy.ts:184), and the test fixture at line 322-350 explicitly returns mcpSessionId to guard against regression.
• tools/call flattening structured content — author confirmed this matches current frontend rendering; revisit when the UI supports rich content.
• Telemetry — pre-existing gap was acknowledged as non-blocking.

Mocking in the new tests is appropriate (mocks at the AWS SDK boundary and at resolveInvokeTarget, which has its own dedicated unit tests in resolve.test.ts).

LGTM.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)